UC San Diego

This dissertation was an investigation into the computational roles of sleep rhythms in the consolidation of memory, and how these roles may be leveraged to the benefit of machine learning and medicine. In Chapter 1 we used an artificial spiking neural network model to validate that a consolidation strategy thought to be taken by the procedural memory system – incrementally learning a new skill by interleaving bouts of training with periods of sleep – can prevent catastrophic forgetting when faced with learning a novel task. In particular, we demonstrated that memory replay during sleep acted to keep the network’s synaptic weight state near to previous memory manifolds as it learns the new task. In Chapter 2, we utilized a biophysical thalamocortical network model to further study this procedural memory consolidation strategy, as well as a declarative memory consolidation strategy – incrementally transferring a new memory to the cortex by hippocampal indexing during sleep. While both strategies were able to prevent catastrophic forgetting, we found that the procedural memory strategy suffers from fine-tuning and works best when training bouts are short and protracted in time. The declarative memory strategy does not suffer from this same fine-tuning problem, suggesting it may be engaged when training bouts are chunked rather than distributed in time. Moreover, our model suggests that the declarative memory consolidation strategy may simply be a compressed version of the procedural memory strategy, with the hippocampus generating simulated training samples to be indexed to the cortex during sleep. We anticipate that such a strategy will be useful in mitigating catastrophic forgetting in machine learning, as others in our lab have shown the procedural memory consolidation strategy to be. Finally, in Chapter 3, we made use of a two-phase biophysical-anatomical and dynamic-neuronal network in order to model the effects of electrical stimulation of the cortical surface and studies the circuit mechanisms behind how this could be used to induce directed traveling waves. We found that cortical surface stimulation differentially recruits distinct subtypes of inhibitory interneurons, which shape the oscillatory frequency and direction of the wave. In the future, we hope to develop this work further to model the induction of sleep rhythms with this network, and how this may be used to aid clinical treatment of memory and sleep disorders.